1
Developing a Large-Scale Tele-Dermatology Program
Session 268, February 14, 2019
Eduardo Cordioli, Telemedicine Medical Manager, Hospital Israelita Albert Einstein
André Santos, Telemedicine IT Project Coordinator, Hospital IsraelitaAlbert Einstein
2
Eduardo Cordioli, MD , MSCE , MBA Telemedicine Medical Manager
André Santos, Telemedicine IT Project Coordinator
Both have no real or apparent conflicts of interest to report.
Conflict of Interest
3
• Brazilian's Physician Demography
• Teledermatology Consultation Flow
• Clinical Staff Management Challenges
• Software Platform: Security, Scalability and Availability
• Deep Neural Networks Training Steps
Agenda
4
• Detail the steps to develop a large-scale social tele-dermatology
program with a low-cost operation
• Identify the most common pitfalls in large database optimization
for medical image acquisition
• Describe how to develop a network of dermatologists available to
screen, diagnose and treat diseases from dermatological images
Learning Objectives
5
Physician demographics in Brazil
• Access to health care is historically a chronic problem in Brazil.
For example, in the 60's the ratio of doctors and inhabitants had a
percentage of only 0.0623%. (SCHEFFER et al, 2018)
• Now in 2018, the percentage is approximately 0.22% However,
the growth does not reflect benefits for the population.
(SCHEFFER et al, 2018)
• There are 452,801 physicians in Brazil. However, the distribution
among regions suffers a great inequality (SCHEFFER et al, 2018).
6
Distribution inequality among
regions
• The percentage of physician distribution
by population varies from 0.28% in the
Southeast to only 0.12% in the North
(SCHEFFER et al, 2018).
• Only the state of São Paulo concentrates
28% of the total number of doctors in the
country (SCHEFFER et al, 2018).
• In contrast, in the state of Maranhão and
Pará, the distribution of physicians by
inhabitants is only 0.09% of the population
(SCHEFFER et al, 2018).
Number of Physicians
per State
7
Challenges providing specialized
care in Brazil
• According to Brazilian Society of Dermatology, from the 5,565 municipalities
only 504 (9.1%) of Brazilian cities have dermatologists.
• Dermatology is the second specialty most referred by primary care
physicians. With a average waiting time of 108 days. (VIEIRA et al, 2014)
36.3
17.78
10.37
10.37
5.93
4.44
14.81
% Of Referral per Speciality
Otolaryngology
Dermatology
Angiology
Orthopedics
Rheumatology
Endocrinology
Others
Speciality
Waiting days
average
Waiting days
maximum
Otolaryngology 399 550
Dermatology 137 108
Angiology 174 174
Orthopedics 68 81
Rheumatology 370 449
Endocrinology 381 507
(VIEIRA et al, 2014)
(VIEIRA et al, 2014)
8
Brazil public health system specialist
consultation flow
3-5 months queue
Patient identify
the symptoms
schedule an appointment
with a general physician
Go to a Basic Health Unit
(UBS)
Get a referral
Receive a phone
call
schedule an appointment
with a specialist
Go to a consultation
With a specialist
Finally starts
treatment
9
Teledermatology consultation flow
1 week diagnosys
Patient identify
the symptoms
schedule an appointment
with a general physician
Go to a Basic Health Unit
(UBS)
APP
Receive Orientation
Maintain Treatment at Basic Unit
70%
Biopsy
3%
Go to a consultation
With a specialist
27%
10
Mobile App
11
Medical Platform
12
Growth Curve
172
2282
3080
3408
4534
6386
11562
12503
1°
Month
2°
Month
3°
Month
4°
Month
5°
Month
6°
Month
7°
Month
8°
Month
Number of diagnosis
13
Teledermatology consultation flow
2a. Phase
Patient identify
the symptoms
Consultation
with a specialist
Starts treatment
GP takes
lesions images
A.I. referrals
the patient
Diagnosys by
Teledermatology
Biopsy procedure
1 week
14
Dermatology Medical Staff – 1st
Reaction
15
Dermatology Medical Staff – 1st
Reactions
• “I do not believe that you could deliver high quality images”
• “We have to see the whole patient”
• “I doubt that general practice doctors or nurses would do the
anamnesis correctly”
• “I already saw a similar attempt in the literature and it did not
work”
• “Are the info and images of the patient in a safe environment"
16
Changing Culture
• Give examples of successful cases .
• Show that they could be paid in the same way as working in
person, with the advantage of doing homeoffice, consulting
dermatology atlas, dictate your own work rhythm.
• Show that the institution's top leadership was committed to the
project.
• Sensitize them by the social scope of the project.
• Possibility of various scientific works.
17
Do the doctor’s trust
Teledermatology?
42%
33%
25%
Confidence in teledermatology diagnosys
before the project
Little confident
Confident
Very confident
17%
83%
Confidence in teledermatology diagnosys
after the project
Confident
Very confident
18
Ok, but… the doctors enjoy
working with Teledermatology?
Would you recommend to a colleague trying work with Teledermatology?
Answer: Yes 100%
Do you want to keep working with Teledermatology?
Answer: Yes 100%
Do you believe that Teledermatology is the future?
Answer: Yes 100%
19
Artificial intelligence in the
clinical operations. Are we there?
20
Microservices architecture
Load
Balancer
Authentication
service
A.I. service
Queue
service
Images Handler
Service
Aplication
service
Report
service
Auto Scaling
21
Information Security
VPN HTTPS
Load
Balancer
Authentication
service
IP
Whitelist
MD5
Encryption
VPN
Concentrator
Local Network
+ WAF
+DDOS
+Firewall
+ Etc
End-To-End Encryption
22
Working With Artificial Intelligence
The first step is to set your goals.
Define what you want to do with your A.I. Algorithm!
1° Give the appropriate referral to the patient
2° Set the priority of care
3° Provide support diagnosys to the clinical staff
23
Clusters optimization
Inflammation of unknown cause
Surface Infection
Eczema
Traumatic cause
Benign tumor
Genetic cause
Pigmentation disturbance
Adverse Drug Reaction
Metabolic cause
Malignant tumor
Benign cyst
Collagenose
Deep infection
Pre-malignant
Psychiatric Disorder
Not clustered
62
30
8
3
10
13
17
6
6
6
22
9
15
8
1
3
210 Possible diagnostics
grouped in 17 cluster
24
Clusters lesions distribution
25
Environment & tools
26
First A.I. Attempt: Original Images
27
• LeNet
• AlexNet
• VGG
• Google Inception
• ResNet
Neural Networks Architectures
28
First A.I. Attempt: Original Images
29
Second A.I. Attempt: Object recognition
30
Third attempt: To develop a system able to
identify the region of interest and crop it
31
Tested Scenarios
Model Train acuracy Validation acuracy
Scenario 1 - ConvNet as fixed feature extractor
Inception V3 Feature Extraction 0.8000 0.6600
VGG19
Feature Extraction
0.4576 0.5000
Resnet50
Feature Extraction
0.2150 0.2000
Scenario 2 - Extending model & re-initialization strategy
Inception V3 Extending Model 0.8893 0.5417
VGG19 Extending Model 0.7124 0.6667
Resnet50 Extending Model 0.9886 0.25
Scenario 3 – Scratch
Inception V3 Scrach 0.9506 0.6562
VGG19
Scrach
0.8417 0.8333
Resnet50
Scrach
0.1946 0.2000
32
Specificity and Sensitivity
Model Specifity Sensitivity
Scenario 1 - ConvNet as fixed feature extractor
Inception V3 Feature Extraction 0.90 0.65
VGG19
Feature Extraction
0.88 0.57
Resnet50
Feature Extraction
0.80 0.20
Scenario 2 - Extending model & re-initialization strategy
Inception V3 Extending Model 0.88 0.56
VGG19 Extending Model 0.91 0.70
Resnet50 Extending Model 0.80 0.20
Scenario 3 – Scratch
Inception V3 Scrach 0.91 0.65
VGG19
Scrach
0.95 0.82
Resnet50
Scrach
0.80 0.20
33
Classification Results
34
• Before training your network, first analyze properly your
database and define your targets with wisdom.
• Do not trust that the network will discover the region of
interest by itself.
• Do not be afraid of run a high number of epochs.
• Always test different algorithms and different learning
paradigms, it will help you to understand your dataset
better.
Lessons learned from CNN training
35
Conclusions
• Brazil has an extremely promising environment for implementation and growth of
distance health care programs due to its size and the concentration of health
professionals in the major economic centers
• The doctors may be resistant at first, but as they start using technology on their daily
routines they start to enjoy it and get more confident every day. This data can be
measured by the growth of doctors answering that they are very confident about
performing virtual diagnosis from 25% to 83%
• Convolutional Neural Networks are perfectly capable of performing tasks such as
screening and diagnostic suggestions when properly designed. The referrals already
reached 83% of acuracy and they have potential to improve even more
• Microservices based architectures and cloud platforms are essential in order to
provide the consistence and availability that a high scaled health operation requires
